AGENCY PLATFORM
The platform an agency runs on. Each person connects their own Claude or ChatGPT to it and gets 200+ tools and the team's knowledge base. About a hundred sites; a new one takes a day, not one to two weeks.
- YEAR
- 2024-2026
- ROLE
- SOLE ENGINEERIN-HOUSE
- STATUS
- IN DAILY USE
STACK
- FRONTEND: Next.js 16 · React 19 · TypeScript · Tailwind 4
- DATA: Supabase · Postgres, self-hosted
- AI: Own MCP server · OAuth · role-based tools · Claude API · OpenAI
- INFRA: Docker · Caddy · VPS · GitHub Actions
- QUALITY: Vitest · Playwright · Lighthouse
CONTEXT
An agency with about a hundred sites. I joined as the developer who built those sites by hand, and each one took one to two weeks. My first fix was a page builder with live preview. Then came AI: for about six months, agents built on OpenClaw ran on the agency's own server, right next to the dashboard.
02 / PROBLEM
That setup was heavy. The agents needed a strong server and someone to keep it running, and the AI bill came to roughly $200 to $300 a month. Meanwhile everyone on the team already had their own Claude subscription, so the same intelligence was being set up a second time on our side.
Sole engineer, and the one who started it: nobody asked for the platform. The dashboard, the hub and its tools, the roles, the database, the background jobs and the servers. Along the way I moved the whole team onto Claude.
MY ROLE
SOLUTION
Move the AI assistants off our server. The dashboard became a hub, and everyone connects their own Claude or ChatGPT to it.
- 01
One hub, 200+ tools. Through it, each person's own assistant works with the sites, the content and the team's knowledge base, the way a colleague would from the dashboard.
- 02
Each role sees only its tools. An editor gets the editing tools, an admin gets more. I choose which tools every person's assistant can reach, so nobody's AI can touch what its owner may not.
- 03
The AI reports back. When an assistant finds a bug or has an idea, it sends a report into the hub and a person sees it in the fix queue. Assistants suggest; people decide and publish.
- 04
A site in a day. A person with their own Claude takes a brief, gets the texts, reviews them, builds the site through the hub and puts it online. One to two weeks became one day.
- 05
Only a small helper stays on the server. Claude Haiku runs the scheduled background jobs and the support chat in the dashboard, and if it does not answer, an OpenAI model takes over.
HOW IT WORKS
A person with their own Claude takes a brief, gets the texts, checks them, builds the site through the hub and puts it online.
WHERE IT STANDS
The agency works in it daily, and the things that mattered stayed true: about a hundred sites, a new one in a day instead of one to two weeks, the blog and Telegram posts on schedule, and one engineer. The change removed a server to look after and a bill to pay.